open-knowledge-studio (oks)

File-based knowledge engineering CLI for Claude Code and coding agents.

oks is the command-line core of Open Knowledge Studio: a file-based knowledge base that turns raw material into a recallable, self-decaying wiki.

Install

pipx install "git+https://github.com/open-agent-power/open-knowledge-studio.git@main#subdirectory=cli" && pipx ensurepath

We recommend pipx because modern Linux (Ubuntu 24.04+) and macOS Homebrew Pythons are PEP 668 externally-managed, so a bare pip install fails. If your mirror lags behind PyPI, add --pip-args="-i https://pypi.org/simple".

Optional multimodal ingest is included in the package, while its heavy dependencies (PDF / audio / video / formula extraction) remain opt-in:

oks capability install watch
oks capability install document
oks capability install pdf

Add --yes to install a listed capability after reviewing its dependencies.

What you get

• 6+1-factor recall engine — token overlap, substring, topic trace, type boost, review penalty, memory curve, plus an optional goal boost that lifts on-scope pages.
• Dreaming cycle — distill raw materials into draft proposals; humans review and promote them to the wiki.
• Decay system — memory-curve scoring with type-specific λ and hot/warm/cold/evictable tiers.
• oks CLI — search, recall, offline evaluation, execution traces, wiki CRUD, drafts, distill, lint, status, metrics.

The CLI core is dependency-light and calls no external network APIs; agents and humans orchestrate the pipeline around it.

Quick start

pipx install "git+https://github.com/open-agent-power/open-knowledge-studio.git@main#subdirectory=cli" && pipx ensurepath   # 1. install the CLI
oks init my-knowledge-base          # 2. scaffold an instance (skills + buckets)
cd my-knowledge-base
oks status                          # 3. use it
oks recall "git branch" --knowledge-only
oks recall "deployment" --knowledge-only --type strategy --format json
oks recall "git branch" --goal none --format json --explain
oks eval recall eval/datasets/team-v1.yaml --output eval/runs/baseline.json
oks trace start memory-goal --run-id demo-001

Use --goal active (default) to merge active goals, --goal <slug> for one reproducible goal, or --goal none for a no-goal baseline. --explain exposes score components without changing ranking.

Machine-readable output uses recall-response/v1 for oks recall and recall-hit/v1 for individual hits. --type filtering happens before ranking and --limit; --knowledge-only drops the episodic path.

oks eval is offline and read-only. oks trace writes append-only execution events under raw/executions/; generated Wiki/Skill proposals stay under drafts/proposals/ until a human explicitly promotes them.

oks init materializes the shareable layer (Claude Code skills, templates, schema, settings) and a git-tracked memory instance. No repo clone required.

Documentation

• Design docs: https://open-agent-power.github.io/open-knowledge-studio/
• Source & issues: https://github.com/open-agent-power/open-knowledge-studio

License

MIT